GS/HW: Avoid detecting shuffles as source of truth for target format

bin/resources/GameIndex.yaml

@@ -12928,7 +12928,6 @@ SCUS-97584:
     autoFlush: 2
     preloadFrameData: 1
     nativeScaling: 2 # Fixes post-processing.
-    halfPixelOffset: 4 # Aligns post bloom.
 SCUS-97589:
   name: "NBA '08 featuring The Life Vol.3"
   region: "NTSC-U"
@@ -21924,7 +21923,6 @@ SLES-52940:
   region: "PAL-M3"
   gsHWFixes:
     halfPixelOffset: 5 # Fixes double image and aligns glow effects.
-    nativeScaling: 2 # Fixes depth blur.
 SLES-52941:
   name: "Gungrave - Overdose"
   region: "PAL-M3"
@@ -45832,7 +45830,6 @@ SLPM-65791:
   region: "NTSC-J"
   gsHWFixes:
     halfPixelOffset: 5 # Fixes double image and aligns glow effects.
-    nativeScaling: 2 # Fixes depth blur.
 SLPM-65793:
   name: "SSX3 [EA BEST HITS]"
   name-sort: "SSX3 [EA BEST HITS]"
@@ -69379,7 +69376,6 @@ SLUS-20969:
   compat: 5
   gsHWFixes:
     halfPixelOffset: 5 # Fixes double image and aligns glow effects.
-    nativeScaling: 2 # Fixes depth blur.
 SLUS-20970:
   name: "Rumble Roses"
   region: "NTSC-U"
@@ -75910,7 +75906,6 @@ SLUS-29132:
   region: "NTSC-U"
   gsHWFixes:
     halfPixelOffset: 5 # Fixes double image and aligns glow effects.
-    nativeScaling: 2 # Fixes depth blur.
 SLUS-29133:
   name: "Namco Transmission Demo Disc Vol.2"
   region: "NTSC-U"
@@ -76713,7 +76708,6 @@ TLES-52940:
   region: "PAL-E"
   gsHWFixes:
     halfPixelOffset: 5 # Fixes double image and aligns glow effects.
-    nativeScaling: 2 # Fixes depth blur.
 TLES-53544:
   name: "Pro Evolution Soccer 5 Beta Trial Code"
   region: "PAL-E"

pcsx2/GS/GSRegs.h

@@ -531,7 +531,6 @@ REG_END2
 	__forceinline bool IsOpaque() const { return ((A == B || (C == 2 && FIX == 0)) && D == 0) || (A == 0 && B == D && C == 2 && FIX == 0x80); }
 	__forceinline bool IsOpaque(int amin, int amax) const { return ((A == B || amax == 0) && D == 0) || (A == 0 && B == D && amin == 0x80 && amax == 0x80); }
 	__forceinline bool IsCd() const { return (A == B) && (D == 1); }
-	__forceinline bool IsAdditive() const { return (A == 0 && B == 2 && (C != 2 || FIX != 0) && D == 1); }
 
 	// output will be Cd, Cs is discarded
 	__forceinline bool IsCdOutput() const { return (C == 2 && D != 1 && FIX == 0x00); }

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -3939,7 +3939,8 @@ void GSRendererHW::Draw()
 
 			if (next_ctx.FRAME.Block() == FRAME_TEX0.TBP0 && next_ctx.FRAME.PSM != FRAME_TEX0.PSM)
 				FRAME_TEX0.PSM = next_ctx.FRAME.PSM;
-			else if (next_ctx.TEX0.TBP0 == FRAME_TEX0.TBP0 && next_ctx.TEX0.PSM != FRAME_TEX0.PSM)
+			// Be careful of the next draw being a channel shuffle!
+			else if (next_ctx.TEX0.TBP0 == FRAME_TEX0.TBP0 && next_ctx.TEX0.PSM != FRAME_TEX0.PSM && GSLocalMemory::m_psm[next_ctx.TEX0.PSM].trbpp >= 16)
 				FRAME_TEX0.PSM = next_ctx.TEX0.PSM;
 			else
 				FRAME_TEX0.PSM = PSMCT32; // Guess full color if no upcoming hint, it'll fix itself later.
@@ -10707,10 +10708,8 @@ int GSRendererHW::IsScalingDraw(GSTextureCache::Source* src, bool no_gaps)
 		IsMipMapDraw() || GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].trbpp <= 8)
 		return 0;
 
-	const bool is_smaller = draw_size.x <= (tex_size.x * 0.75f) && draw_size.y <= (tex_size.y * 0.75f);
-	const bool is_smaller_and_writing_back = draw_size.x < tex_size.x && draw_size.y < tex_size.y && next_ctx.FRAME.Block() == m_cached_ctx.TEX0.TBP0 && next_ctx.TEX0.TBP0 == m_cached_ctx.FRAME.Block();
 	// Should usually be 2x but some games like Monster House goes from 512x448 -> 128x128
-	const bool is_downscale = m_cached_ctx.TEX0.TBW >= m_cached_ctx.FRAME.FBW && (is_smaller || is_smaller_and_writing_back);
+	const bool is_downscale = m_cached_ctx.TEX0.TBW >= m_cached_ctx.FRAME.FBW && draw_size.x <= (tex_size.x * 0.75f) && draw_size.y <= (tex_size.y * 0.75f);
 	// Check we're getting most of the texture and not just stenciling a part of it.
 	// Only allow non-bilineared downscales if it's most of the target (misdetections of shadows in Naruto, Transformers etc), otherwise it's fine.
 	const GSVector4i src_valid = src->m_from_target ? src->m_from_target->m_valid : src->m_valid_rect;
@@ -10734,8 +10733,7 @@ int GSRendererHW::IsScalingDraw(GSTextureCache::Source* src, bool no_gaps)
 	}
 
 	// Last ditched check if it's doing a lot of small draws exactly the same which could be recursive lighting bloom.
-	if (m_vt.m_primclass == GS_SPRITE_CLASS && m_index->tail > 2 && !no_gaps_or_single_sprite && m_context->TEX1.MMAG == 1 && is_target_src && (m_cached_ctx.TEST.ZTST == ZTST_ALWAYS || m_cached_ctx.TEST.ZTE == 0) &&
-		!m_context->ALPHA.IsOpaque() && GSLocalMemory::m_psm[m_cached_ctx.FRAME.PSM].bpp == GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM].bpp)
+	if (m_vt.m_primclass == GS_SPRITE_CLASS && m_index->tail > 2 && !no_gaps_or_single_sprite && m_context->TEX1.MMAG == 1 && !m_context->ALPHA.IsOpaque())
 	{
 		GSVertex* v = &m_vertex->buff[0];
 		float tw = 1 << src->m_TEX0.TW;
@@ -10746,10 +10744,7 @@ int GSRendererHW::IsScalingDraw(GSTextureCache::Source* src, bool no_gaps)
 		const int first_x = (v[1].XYZ.X - v[0].XYZ.X) >> 4;
 		const int first_y = (v[1].XYZ.Y - v[0].XYZ.Y) >> 4;
 
-		const u32 half_color_mask = (m_vt.m_min.c == GSVector4i(64, 64, 64, 64)).mask() & 0xfff;
-		const bool possible_fake_bilinear = no_resize && PRIM->ABE && m_context->ALPHA.IsAdditive() && m_cached_ctx.TEX0.TFX == TFX_MODULATE && m_vt.m_eq.rgba && half_color_mask == 0xfff;
-
-		if (first_x >= first_u && first_y >= first_v && (!no_resize || possible_fake_bilinear) && ((draw_size.x != first_x && draw_size.y != first_y) || (tex_size.x != first_u && tex_size.y != first_v)) && std::abs(draw_size.x - first_x) <= 4 && std::abs(draw_size.y - first_y) <= 4)
+		if (first_x > first_u && first_y > first_v && !no_resize && std::abs(draw_size.x - first_x) <= 4 && std::abs(draw_size.y - first_y) <= 4)
 		{
 			for (u32 i = 2; i < m_index->tail; i += 2)
 			{
@@ -10758,14 +10753,14 @@ int GSRendererHW::IsScalingDraw(GSTextureCache::Source* src, bool no_gaps)
 				const int next_x = (v[i + 1].XYZ.X - v[i].XYZ.X) >> 4;
 				const int next_y = (v[i + 1].XYZ.Y - v[i].XYZ.Y) >> 4;
 
-				if (std::abs(draw_size.x - next_x) > 4 || std::abs(draw_size.y - next_y) > 4 || (draw_size.x == next_x && tex_size.x == next_u) || (draw_size.y == next_y && tex_size.y == next_v))
+				if (std::abs(draw_size.x - next_x) > 4 || std::abs(draw_size.y - next_y) > 4)
 					break;
 
 				if (next_u != first_u || next_v != first_v || next_x != first_x || next_y != first_y)
 					break;
 
 				if (i + 2 >= m_index->tail)
-					return (is_target_src && src->m_from_target->m_downscaled) ? -1 : 2;
+					return 2;
 			}
 		}
 	}

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -2580,12 +2580,21 @@ GSTextureCache::Target* GSTextureCache::LookupDrawTarget(GIFRegTEX0 TEX0, const
 			{
 				// Some games misuse the scissor so it ends up valid 1 pixel over, which causes hell for us. So check if it still overlaps without the extra pixel.
 				const GSVector4i adjusted_valid = GSVector4i(t->m_valid.x, t->m_valid.y, std::min(t->m_valid.z, static_cast<int>(t->m_TEX0.TBW) * 64), t->m_valid.w - 1);
-				const u32 adjusted_endblock = GSLocalMemory::GetEndBlockAddress(t->m_TEX0.TBP0, t->m_TEX0.TBW, t->m_TEX0.PSM, adjusted_valid);
+				u32 adjusted_endblock = GSLocalMemory::GetUnwrappedEndBlockAddress(t->m_TEX0.TBP0, t->m_TEX0.TBW, t->m_TEX0.PSM, adjusted_valid);
 				if (adjusted_endblock <= bp)
 				{
 					i++;
 					continue;
 				}
+				const GSVector4i adjusted_rect = GSVector4i(min_rect.x, min_rect.y, std::min(min_rect.z, static_cast<int>(t->m_TEX0.TBW) * 64), min_rect.w - 1);
+				// Also check the offset from the bp to the current request to see if it overlaps the begining of the selected target.
+				adjusted_endblock = GSLocalMemory::GetUnwrappedEndBlockAddress(TEX0.TBP0, TEX0.TBW, TEX0.PSM, adjusted_rect);
+				if (adjusted_endblock <= t->m_TEX0.TBP0)
+				{
+					i++;
+					continue;
+				}
+
 				const GSLocalMemory::psm_t& s_psm = GSLocalMemory::m_psm[TEX0.PSM];
 				const u32 widthpage_offset = (std::abs(static_cast<int>(bp - t->m_TEX0.TBP0)) >> 5) % std::max(t->m_TEX0.TBW, 1U);
 				const bool is_aligned_ok = widthpage_offset == 0 || ((min_rect.width() <= static_cast<int>((t->m_TEX0.TBW - widthpage_offset) * 64) && (t->m_TEX0.TBW == TEX0.TBW || TEX0.TBW == 1)) && bp >= t->m_TEX0.TBP0);